Keyword Intelligence Subnet
Keyword Intelligence Subnet is a decentralized SEO intelligence network on Bittensor
视频
技术栈
描述
Links:
Social Post: https://x.com/OzanAndac_/status/2026686589376643534
Proposal: https://www.notion.so/Keyword-Intelligence-Subnet-31160c83848b80968f85da9a139e4013?source=copy_link
Introduction
Keyword Intelligence Subnet (name subject to change) is a decentralized SEO intelligence network on Bittensor that uses competitively scored miners to estimate search volume, CPM, keyword difficulty, and target keywords for websites through consensus-driven statistical modeling.
Miners are tasked with analyzing websites and web pages to extract keyword data, including volume, cost-per-mille (CPM), keyword difficulty, and relevance scores. Miners can support different types of data without a penalty, since SEO data is broad, and not everybody can have access to everything. The goal is to let miners do what they are best at.
Validators score each data type with a different approach, and to be sure that an average metric is returned for a specific keyword.
Problem & Solution
If you dealt with blogging or local SEO before, you are likely to use a keyword analysis tool to analyze your target keyword and monitor your competitors.
Tools like SEMrush, Ahrefs, and Moz charge high monthly fees because they operate massive centralized scraping and clickstream infrastructures. Small startups, indie founders, and agencies are priced out. Most individuals are exploiting the freemium plan of these tools, which single-handedly shows the fact that we lack a public and affordable tool.
That’s why KIS can build a new platform using miners with different skills. One miner can analyze keywords while another can calculate a TA score.
Our collaboration against their centralized clickstream data.
Incentive & Mechanism Design
KIS operates as a competitive estimation market rather than a shared scraping pool. Miners are not forced to implement the same methodology. Instead, they compete on the quality, stability, and speed of their keyword estimations.
Each task type is treated independently at the validator level. A miner may specialize in one or several metrics. Performance is tracked separately per metric so that a strong CPM estimator is not penalized for not supporting keyword volume estimation.
However, if a miner is specialized at one task and returns an inaccurate estimate, it will get penalized.
When a validator receives a keyword query, it sends the same task to multiple miners. Responses are aggregated using statistical filtering. Extreme outliers are removed. A consensus value is computed using a median or trimmed mean. Each miner is then scored based on how close its estimate is to that consensus value. Over time, validators maintain historical performance records to detect instability, randomness, or manipulation patterns.
Speed is also factored into scoring, but it is weighted carefully to avoid incentivizing low-quality fast responses. Stability across repeated queries matters more than one-off precision. A miner that fluctuates heavily for identical inputs will gradually lose weight.
Miners
Primary purpose of a KIS miner is returning keywords from a web page, and a TA score as an output. The rest of the tasks are extra points for a miner because, without finding the primary keyword of a web page or a website, there is no point in analyzing it.
An example request miner receives would be a JSON data like this:
{
"page": URL(String)
}
And the minimum data a miner can return is:
{
"primary_keyword": String,
"LSI": String[],
"ta_score": Number
}
A primary keyword is the main focus of the web page. LSIs are alternative keywords that help search engines understand the context of the page.
To determine the correctness of keywords, we need to compare the miner results with each other. This is a task of validators, which is explained in the next section.
If a miner owner has access to tools that help them to calculate keyword metrics (like paid keyword analysis tools, Google Ads, or keyword databases like DataForSEO provides…), they can increase their weights by providing extra information.
Even if a miner has access to advanced tools, they don’t have to provide everything at once. The most general data a miner can return as an output is:
{
"primary_keyword": String,
"LSI": String[],
"ta_score": Number
"region": String,
"relevance_score": Number,
"volume": Number,
"cpm": Number,
"difficulty": Number,
"intent": String
}
This type of data gives enough information about a keyword.
What is TA score?
A Tensor Authority (TA) score is an alternative metric to evaluate the “worth” of a web page.
Unlike traditional authority metrics that rely on proprietary backlink indexes and undisclosed machine learning models, TA is a consensus-driven authority estimate produced by competitively scored miners within the subnet.
TA measures how strong a page is in terms of its ability to answer. TA does not measure backlink strength alone. Instead, it measures how well a page actually satisfies search intent and delivers useful content for the target query.
For a given keyword–page pair, miners evaluate signals such as:
Whether the page directly answers the intended query
Depth and completeness of explanation
Topical relevance and semantic coverage
Presence of spam patterns or keyword stuffing
Content structure and clarity
Use of authoritative references
Readability and information density
Each miner applies its own content-quality and intent-matching model to produce a normalized TA score between 0 and 100.
Validators then collect multiple TA estimates, remove statistical outliers, and compute a consensus score. Miners are rewarded based on how closely their assessment aligns with statistically reinforced consensus and long-term stability.
It is an alternative metric to DA (Domain Authority) and PA (Page Authority). We collect this information so that in the future, if SEO is mostly replaced with AEO, we can be ready for new ranking metrics. KIS has a sole purpose of collecting TA data and storing it for the future.
Validators
Validators evaluate the quality of miner responses using metric-specific scoring methods. Numerical outputs such as volume or CPM are assessed based on their deviation from the consensus average, while keyword outputs are validated using semantic similarity and word-matching algorithms.
Validating Keyword Correctness
Miners are have to return a primary keyword as well as at least one LSI keyword in an array. To determine if a primary keyword fits the overall web page, we need to find similarities between miner outputs and punish those that return inaccurate results.
A simple but not the best approach would be to use Sentence Transformers and cosine similarity.
from sentence_transformers import SentenceTransformer
from sklearn.metrics.pairwise import cosine_similarity
import numpy as np
import re
def clean_text(text):
text = text.lower()
text = re.sub(r"[^\\w\\s]", "", text)
return text.strip()
def token_overlap_score(primary, candidate):
p_tokens = set(clean_text(primary).split())
c_tokens = set(clean_text(candidate).split())
if not p_tokens:
return 0
overlap = len(p_tokens.intersection(c_tokens))
return overlap / len(p_tokens)
def score_keywords(primary_keyword, keyword_list):
model = SentenceTransformer(MODEL_NAME)
# Encode primary once
primary_embedding = model.encode([primary_keyword])
# Batch encode candidates
candidate_embeddings = model.encode(keyword_list)
# Cosine similarity
similarities = cosine_similarity(primary_embedding, candidate_embeddings)[0]
results = []
for i, keyword in enumerate(keyword_list):
semantic_score = (similarities[i] + 1) / 2 * 100 # normalize 0-100
# Exact match bonus
exact_bonus = EXACT_MATCH_BONUS if clean_text(primary_keyword) == clean_text(keyword) else 0
# Token overlap score (0-100 scale)
overlap_score = token_overlap_score(primary_keyword, keyword) * 100
# Final hybrid score
final_score = (
SEMANTIC_WEIGHT * semantic_score +
TOKEN_OVERLAP_WEIGHT * overlap_score +
exact_bonus
)
final_score = min(round(final_score, 2), 100)
results.append({
"keyword": keyword,
"score": final_score
})
# Sort descending
results.sort(key=lambda x: x["score"], reverse=True)
return results
Even though that’s not the best approach, it visualizes how a validator would compare different keyword results well enough.
Business Logic & Market Rationale
KIS aims to build an infrastructure layer for keyword and intent intelligence, not just another SEO dashboard.
Traditional SEO tools operate as closed systems. They collect data, process it internally, and sell access to their database. Users have no control over how data is calculated and no way to improve the system. Everything depends on one company.
KIS works differently. It separates data production from data consumption.
Miners produce intelligence. Validators verify it. The platform aggregates it.
This structure makes the system scalable. Instead of building one massive centralized crawler, we distribute the workload across many independent miners. Instead of trusting one model, we rely on statistical consensus.
From a market perspective, SEO continues to grow. Content creators, agencies, e-commerce brands, SaaS startups, and local businesses all depend on keyword data. Meanwhile, the cost of existing tools keeps rising.
KIS positions itself as:
A lower-cost alternative
A transparent scoring system
A decentralized keyword estimation layer
A future-ready infrastructure for AEO and AI search
The long-term value lies not only in selling subscriptions, but in building a public dataset of keyword intelligence and TA scores that can be reused in other products.
If AI search replaces traditional SEO, structured intent and helpfulness data will become even more important. KIS is preparing for that shift.
Go-To-Market Strategy
Building a SEMrush alternative isn’t easy. First, we need a working prototype and start collecting data.
Roadmap
1) A working Subnet
At first phase, we need a working prototype. Until multiple miners start analyzing each data type of keyword metrics we can’t talk about an MVP product.
That’s why Subnet is the priority and the starting point.
2) Data Analysis
SEMrush and leader keyword analysis tools use machine learning to calculate keyword volumes.
Here’s what the process of volume calculation looks like:
SEMrush takes sets of data from third-party providers and overlays it with historical clickstream data acquired from reliable sources.
SEMrush forwards the resulting set of data to a computing cluster and receives a computational model.
SEMrush applies this model to get the volumes for billions of keywords.
https://www.semrush.com/kb/683-what-is-search-volume-in-semrush
While validators and miners collect and validate data, we will need another system & subnet to analyze data using machine learning.
3) Product Launch
Now we are ready to talk about an advanced SEO & keyword analysis tool that can compete with big players. It’s time to build the actual product on top of actively running Subnet and historical Tensor Authority data.
本次黑客松进展
Proposal is completed
融资状态
Self-funded